NeurIPS 2025 Papers

Out-of-distribution (OOD) detection is essential for reliably deploying machine learning models in the wild. Yet, most methods treat large pre-trained models as monolithic encoders and rely solely on their final-layer representations for detection. We challenge this wisdom. We reveal the intermediate layers of pre-trained models, shaped by residual connections that subtly transform input projections, can encode surprisingly rich and diverse signals for detecting distributional shifts. Importantly, to exploit latent representation diversity across layers, we introduce an entropy-based criterion to automatically identify layers offering the most complementary information in a training-free setting, without access to OOD data. We show that selectively incorporating these intermediate representations can increase the accuracy of OOD detection by up to $10\%$ in far-OOD and over $7\%$ in near-OOD benchmarks compared to state-of-the-art training-free methods across various model architectures and training objectives. Our findings reveal a new avenue for OOD detection research and uncover the impact of various training objectives and model architectures on confidence-based OOD detection methods.

Contrastive Language-Image Pretraining (CLIP) has demonstrated strong zero-shot performance across diverse downstream text-image tasks. Existing CLIP methods typically optimize a contrastive objective using negative samples drawn from each minibatch. To achieve robust representation learning, these methods require extremely large batch sizes and escalate computational demands to hundreds or even thousands of GPUs. Prior approaches to mitigate this issue often compromise downstream performance, prolong training duration, or face scalability challenges with very large datasets. To overcome these limitations, we propose AmorLIP, an efficient CLIP pretraining framework that amortizes expensive computations involved in contrastive learning through lightweight neural networks, which substantially improves training efficiency and performance. Leveraging insights from a spectral factorization of energy-based models, we introduce novel amortization objectives along with practical techniques to improve training stability. Extensive experiments across 38 downstream tasks demonstrate the superior zero-shot classification and retrieval capabilities of AmorLIP, consistently outperforming standard CLIP baselines with substantial relative improvements of up to 12.24%.

Standard training metrics like loss fail to explain the emergence of complex capabilities in large language models. We take a spectral approach to investigate the geometry of learned representations across pretraining and post-training, measuring effective rank (RankMe) and eigenspectrum decay (αReQ). With OLMo (1B-7B) and Pythia (160M-12B) models, we uncover a consistent non-monotonic sequence of three geometric phases during autoregressive pretraining. The initial “warmup” phase exhibits rapid representational collapse. This is followed by an “entropy-seeking” phase, where the manifold’s dimensionality expands substantially, coinciding with peak n-gram memorization. Subsequently, a “compression-seeking” phase imposes anisotropic consolidation, selectively preserving variance along dominant eigendirections while contracting others, a transition marked with significant improvement in downstream task performance. We show these phases can emerge from a fundamental interplay of cross-entropy optimization under skewed token frequencies and representational bottlenecks (d ≪ |V|). Post-training further transforms geometry: SFT and DPO drive “entropy-seeking” dynamics to integrate specific instructional or preferential data, improving in-distribution performance while degrading out-of-distribution robustness. Conversely, RLVR induces “compression-seeking” , enhancing reward alignment but reducing generation diversity.

Foundation models pretrained on web-scale data drive contemporary transfer learning in vision, language, and multimodal tasks. Recent work shows that mild label noise in these corpora may lift in-distribution accuracy yet sharply reduce out-of-distribution generalization, an effect known as catastrophic inheritance. Medical data is especially sensitive because annotations are scarce, domain shifts are large, and pretraining sources are noisy. We present the first systematic analysis of catastrophic inheritance in medical models. Controlled label-corruption experiments expose a clear structural collapse: as noise rises, the skewness and kurtosis of feature and logit distributions decline, signaling a flattened representation space and diminished discriminative detail. These higher-order statistics form a compact, interpretable marker of degradation in fine-grained tasks such as histopathology. Guided by this finding, we introduce a fine-tuning objective that restores skewness and kurtosis through two scalar regularizers added to the task loss. The method leaves the backbone unchanged and incurs negligible overhead. Tests on PLIP models trained with Twitter pathology images, as well as other large-scale vision and language backbones, show consistent gains in robustness and cross-domain accuracy under varied noise levels.

Machine learning models play a key role in safety-critical applications, such as autonomous vehicles and advanced driver assistance systems, where their robustness during inference is essential to ensure reliable operation. Sensor faults, however, can corrupt input signals, potentially leading to severe model failures that compromise reliability. In this context, pretraining emerges as a powerful approach for learning expressive representations applicable to various downstream tasks. Among existing techniques, masking represents a promising direction for learning representations that are robust to corrupted input data. In this work, we extend this concept by specifically targeting robustness to sensor outages during pretraining. We propose a self-supervised masking scheme that simulates common sensor failures and explicitly trains the model to recover the original signal. We demonstrate that the resulting representations significantly improve the robustness of predictions to seen and unseen sensor failures on a vehicle dynamics dataset, maintaining strong downstream performance under both nominal and various fault conditions. As a practical application, we deploy the method on a modified Lexus LC 500 and show that the pretrained model successfully operates as a substitute for a physical sensor in a closed-loop control system. In this autonomous racing application, a supervised baseline trained without sensor failures may cause the vehicle to leave the track. In contrast, a model trained using the proposed masking scheme enables reliable racing performance in the presence of sensor failures.